# OpenAI and Broadcom Unveil a Custom Inference Chip Named Jalapeño

A purpose-built accelerator for serving large language models shows another large buyer designing custom hardware to reduce its reliance on Nvidia for inference.

- Published: 2026-06-25T10:46:42.427Z
- Canonical: https://polylog.news/ai/2026-06-25/openai-and-broadcom-unveil-a-custom-inference-chip-named-jal
- Publisher: Polylog (AI desk)
- Section: markets
- Sources: [OpenAI News](https://openai.com/index/openai-broadcom-jalapeno-inference-chip), [Nvidia Blog](https://blogs.nvidia.com/blog/nvidia-aws-ai-production-scale/)

OpenAI and Broadcom have introduced a custom accelerator, called Jalapeño, built specifically for large language model inference rather than training, [according to OpenAI](https://openai.com/index/openai-broadcom-jalapeno-inference-chip). The case for inference-tuned chips is the standard one: better performance per dollar and per watt on the serving workloads that now dominate a frontier lab's compute costs, since a model is trained once but served continuously.

OpenAI joins a pattern already set by Google's tensor processing units and Amazon's Trainium and Inferentia lines. The economics are straightforward. Inference is a large, predictable, and growing cost, and a buyer at OpenAI's scale can spread the considerable expense of a custom design across enough chips to undercut merchant graphics processing units (GPUs) on its own traffic. Broadcom, which designs the application-specific integrated circuits behind several hyperscaler programs, is the supplier that benefits whether or not any single lab's chip succeeds.

The same day, Nvidia and Amazon Web Services published [joint work on production-scale inference](https://blogs.nvidia.com/blog/nvidia-aws-ai-production-scale/) covering low-latency serving and GPU price-performance, a reminder that the incumbent is defending the serving layer that custom silicon targets most directly. OpenAI did not disclose the relevant numbers, throughput, latency, and cost per million tokens against a current Nvidia baseline, so the size of any advantage is asserted rather than demonstrated.

## What this means

Every large model buyer that designs its own inference chip reduces Nvidia's share of the fastest-growing part of AI spending and shifts value toward custom-silicon designers like Broadcom. The threat is concentrated in inference, where workloads are stable enough to justify a fixed-function part, and far weaker in training, where flexibility still favors general GPUs.

## What to watch

- Independent benchmarks of Jalapeño's cost and latency against a current Nvidia part, since vendor-stated efficiency gains routinely shrink under third-party testing.
- Whether OpenAI shifts a meaningful share of its own serving traffic onto the chip, the only real proof that a custom design is more economical than buying GPUs.